NVIDIA Unveils Cosmos Framework, NIM Microservices, and Hardware Specifications for Physical AI

NVIDIA has released detailed implementation recipes and GPU requirements for its Cosmos world models, targeting autonomous vehicle, robotics, and surgical simulation workflows. The announcements introduce the open Cosmos Framework and NVIDIA Inference Microservices (NIM) across more than 20 supported GPUs. Cosmos NIM delivers 2x to 3x mean inference speedups on H200 GPUs via speculative decoding, along with sub-second streaming VLM alerts. For deployments, Cosmos3-Nano requires 31 GiB to 58 GiB of memory, while Cosmos3-Super requires 38 GiB to 150 GiB depending on precision and offloading configurations.